Graphically representing child-robot interaction proxemics

Marie D. Manner1, Maria Gini1, and Jed Elison2

1Department of Computer Science and Engineering2Institute of Child Development

University of Minnesotamanne044, gini, jtelison @umn.edu

Abstract

This paper discusses the current analysis of a large set ofchild-robot interaction experiments, in which a 2–4 year oldchild plays a set of games with a small humanoid robot, withthe goal of detecting early signs of autism in toddlers based onthe interactions. Our first goal in this paper is to condense anddisplay these child-robot interactions as multi-channel timeseries, starting with the distances between the child, robot,parent, and experimenter. Our second goal is to use these datadisplays to compare and contrast different children, with theaim of clustering children with similar interaction patterns.Using a ceiling-mounted camera, we record the interactionbetween a child and a robot which performs different gamesand dances. After analyzing the video footage for the loca-tions of all people and the robot in the room over the variablelength of the interaction, we condense the interactions intosimplified, quantifiable, scale-invariant data. We show howthe distances between the child and robot, experimenter, andcaregiver can be discretized into a few location zones andcompared across children using classic similarity measures.Proxemics (social distances) between the child, robot, care-giver, and experimenter during a child-robot interaction showdifferences between participants and hence can provide addi-tional information for behavior analysis.

1 IntroductionOur work is part of a larger experimental project in which weseek to automate the analysis of the reaction of a child whiles/he plays games with a robot. This paper shows how wecondense and display the movements of each child with re-spect to the robot, the child’s caregiver, and the experimenterthroughout the interaction, and compare all participants toeach other. The overarching goal is to assist in the detec-tion of abnormal development by leveraging the interest ofchildren with autism in robots, as established in (Scassellati2007). The work shown here is with neurotypical children,giving us a baseline with which to compare future partic-ipants. Symptoms of some pervasive developmental disor-ders, such as autism spectrum disorder (ASD), include dif-ferences in personal space between individuals and objectsor people (Asada et al. 2016), eye contact, physical contact,and a longer delay or non-response when called by name,

Copyright c© 2018, Association for the Advancement of ArtificialIntelligence (www.aaai.org). All rights reserved.

Figure 1: Sample overhead view of the human-robot inter-action experiment.

among other differences (Zwaigenbaum et al. 2005). Thesebehaviors are thus considered early markers for autism. Wepropose methods to generate a portable, reproducible, au-tomated behavior analysis that can help detect these earlymarkers.

The children we work with may be on the autism spec-trum or have another pervasive developmental disorder, sowe want to avoid adding fiducial markers to the child whilemaintaining a structured and reproducible play interaction.We thus record all interactions from several different per-spectives; the only perspective that can reliably capture theparticipant’s location at all times is from the ceiling, so wemount a camera on the ceiling in the center of the room. Fig.1 shows this vantage point, taken from a GoPro; distortionexists but still allows for identifying an individual’s positionover time.

The work we present here has two goals, framed as ques-tions. First, how can we compactly represent the interactionof a child and robot as a time series? Second, how can wecompare participants to each other on the basis of the inter-action? By answering these questions, we provide a method

to cluster participants based solely on proxemics throughoutthe interaction. The results for our neurotypical participantswill provide a baseline with which to compare results fromchildren at high risk for autism or even children with clin-ical diagnoses. The responses of high risk children, whencompared to these neurotypical participants, could end upbeing outliers or could be similar to an identifiable subsetof children reactions. In these cases, the results will providea potentially clinically useful identifier, which could then bematched to other data collected on the participants and addedto the autism phenotype. If instead the responses of high riskchildren look very similar to neurotypical children, we willknow that children in this age group behave similarly forthese interactions. In this case, future research will need tocarefully consider how autonomous behaviors of robots af-fect reaction for this age group.

This paper contributes several ways of organizing and dis-playing child-robot interactions as time series data. We showhow difference scores between interactions can be displayedgraphically, how individual interactions can be displayed asinteraction zones, how interaction zones simplify the visu-alization of how child comfort levels change over time, andhow the interaction zones compare across children. We be-gin with related work in Section 2, give a high-level descrip-tion of the human-robot interaction experiment and raw datataken therefrom in Section 3, describe the preliminary anal-ysis in Section 4, and end with a summary in Section 5.

2 Related WorkSocially Assistive Robotics is a recent area of robotics re-search aimed at helping populations with special needs; itincludes research for children or adults, such as robots astools for children with pervasive developmental disordersor robots for adults as tools, companions, or helpers. So-cially assistive robotics research in autism is over a decadeold, yet does not currently meet standards of psychol-ogy and child development researchers (Diehl et al. 2012;Scassellati, Mataric, and Admoni 2012; Pennisi et al. 2016).Robotics research with children with autism stems fromthe fact that afflicted children tend to especially enjoy au-tonomous (or seemingly autonomous) robots (Dautenhahnand Werry 2004), and researchers have used a wide varietyof robot appearances and abilities in this area (Scassellati,Mataric, and Admoni 2012). While the reason for this highlevel of interest is unknown, researchers clearly have thepotential to leverage robotics for autism diagnosis or treat-ment (Scassellati 2005). We now give a brief overview ofrelated work for this paper, detecting autism traits throughautomated video analysis and proxemics.

Automatically detecting autism or autistic traits is a cur-rent research area in computer vision, and much work usesas much data as possible. For example, Hashemi et al. (2012)analyzed non-intrusive camera footage using a GoPro placedon a table, two to four feet from a clinician-child pair inwhich the clinician was testing the child with a disengage-ment of attention task and a visual tracking task. The au-thors went even further in (Hashemi et al. 2014), in whichthey analyzed interest sharing and atypical motor behaviorby estimating head motions from facial features and motor

behavior by arm symmetry. Fasching et al. (2015) automat-ically coded activities of people with obsessive-compulsivedisorders from overhead video footage in a structured lab,tracking how many times participants touched various ob-jects. These objects are statically located, such as faucetsand handles, and easier to locate in a static environment. Incontrast, our laboratory works with very young children ina play-based interaction, which adds difficulties in instru-menting the room and reliably tracking an active, potentiallynon-cooperative child.

Much research in socially assistive robotics studies child-robot interactions. Feil-Seifer and Mataric (2011) created ashort free-play interaction with children and a robot, withthe future intention of allowing a robot to adjust its own be-havior based on the child’s reaction. This work tracked thechild in relation to the robot to automatically determine ifthe child was having a positive or negative reaction to therobot. The authors manually coded for the child avoidingthe robot, interacting with the robot or playing with bubblesthe robot generated, staying still, being near parent, beingagainst the wall, or none of those. Results showed that chil-dren with a positive reaction to the robot spent over 80%of time interacting, whereas children with a negative reac-tion spent less than 20% of time interaction with the robot.Mead et al. (2013) also investigated proxemics (the studyof social distances), by placing a participant and researcherin discussion about a static humanoid robot. Using a videocamera and depth data, they studied body pose during theexperiment, training Hidden Markov Models on sensory ex-periences (such as voice loudness and a variety of distancesto other people and environment objects) to correctly anno-tate initiation and termination of conversation.

3 Research Method3.1 Experimental ParadigmThe overarching goal of the robot interaction study withtoddlers (the age group of roughly 2–3 years old) is toidentify children at high risk for autism spectrum disor-der (ASD). Children were recruited from a laboratory-maintained database at the University of Minnesota’s Insti-tute of Child Development. The participants in this studyare very low to medium risk for ASD, but mostly low risk.Thus, the data in this paper represents a mostly neurotyp-ical set of children, providing us with a reactionary base-line for future comparisons with very high risk or diagnosedtoddlers. Written informed parental consent was ensured inadvance of all testing; all research was approved by the uni-versity’s Institutional Review Board. We collected multipledata sets from each participant, including standardized andnovel assessments and video footage, but for this paper wereview and discuss only video footage taken from an over-head perspective through-out the duration of the interactionwith the child. We wrote software to track the actors of inter-est throughout the video, usually the child, the robot, exper-imenters, and/or caregivers, and the analysis presented hereis generated from the raw coordinates of each actor in everyframe of the video.

During the experiments, we introduce the child to Robbie

the Robot (a NAO from Softbank Robotics). Robbie playsdifferent games such as looking games, imitation games, anddances. The games include “I Spy” (a looking game that en-courages the child to find objects in the room), “Simon Says”(a behavior imitation game that encourages the child to copymotions possible with gross motor skills like clapping andwaving), and several dances set to music. The set of gamesis in the same order for every child. The experimenter con-trolling the robot imitates some of the robot’s movementsand plays along during some of the looking games, encour-aging the child to do the same. The interaction is recordedfrom up to four perspectives, most notably from a GoPromounted on the ceiling. The GoPro is the only camera thatis always located in the same place, is impossible to reachby participants, and from which we can almost always seeall actors in the room.

We attempted to keep the same location for all actors inthe room across experiments. The experimenter that con-trolled the robot (hereafter called simply the experimenter)sat next to the robot slightly off-center in the room. If thechild did not need comfort or attention from their caregiver,they were seated or standing on the floor near the robot, fac-ing the robot. In this case, the caregiver sat near the edge ofthe room with another researcher, answering questions froma development assessment. If the child needed constant orfrequent attention from the parent, the child might be seatedon or near their parent during all or part of the interaction,usually closer to the robot than the parent would be if thechild did not need attention. Fig. 1 shows part of an over-head video frame; the child faces the robot, seated on theircaregiver, and the experimenter is next to the robot.

3.2 DataIn all, 65 participants were recruited for this study, of which60 contributed video footage of a robot interaction. In onecase, the experimenter, a second researcher, and the parentall sat on the floor and attempted to draw the child’s attentionto the robot; as this was a highly unusual configuration, thisvideo was not analyzed for this paper, leaving us with 59videos. These 59 participants, 31 males, 28 females, wereaged 25 to 45 months with a mean of 32.9 months and stan-dard deviation of 4.6 months. The interactions and thereforevideos range from roughly nine to 15 minutes long, depend-ing on the child’s willingness or ability to continue interact-ing with the robot. The original video is slightly distorted,thus we first perform an un-distortion and use the undistortedvideo for later analysis.

There are seven total parts to the interaction (games anddances), which we call presses for attention, or presses. Eachpress varies in time, and there is a one minute buffer betweeneach press to give the child time to re-engage if they werenot interested or took a break from playing for some reason(e.g. requested a snack). If the experimenter did not triggerthe next press, after one minute passed the robot started thenext press anyway. Some interactions were not completeddue to equipment malfunction or participant choice, givingus some children who did not complete all seven presses.Where possible, these data are included in the analyses.

The data used here stems from the raw coordinates of each

actor in the room. First, we find the Euclidean distance be-tween actors, giving us three channels of data: the distancesbetween the child and robot, the child and caregiver, andthe child and experimenter. In choosing Euclidean distancesbetween actors over time, we start on our first goal, to rep-resent an interaction over time. An example of this data isshown in Fig. 2; the distance between the child and care-giver, shown in a solid green line, starts at 0 feet, which isalso the frame shown in Fig. 1. Around minute 4 of the in-teraction, the child began to move away from the caregiver,shown in Fig. 3, and away from the caregiver, robot, and ex-perimenter around minute 6 of the interaction, shown in Fig.4. For sake of continuity, most of the graphs or figures in thispaper that show a single child participant’s data comes fromthe same child; the only exception is in Fig. 11.

Figure 2: The Euclidean distances between the child and par-ent, child and robot, and child and experimenter (smoothedby averaging over every second). Blue vertical bars indicatethe beginning of a press for social interaction.

Figure 3: The child movesaway from caregiver aroundminute 4.

Figure 4: The child movesacross the room aroundminute 6.

In some experiments, two caregivers were present dur-ing the interaction; the minimum distance between the childand either parent was used in our data, ensuring that wecan reasonably compare children with one or two care-givers present. The interactions, which are recorded at 33

frames/second, were reduced by averaging the Euclideandistances in one second windows to smooth the data slightly;this averaged dataset is what we used for data analysis.

4 Data Analysis and ResultsOur first data reduction method reduces the three time seriesper child into one, essentially normalizing the data.

Figure 5: Condensing three time series to a single metric,with and without considering the distance to the experi-menter (smoothed by averaging over every second). Bluevertical bars indicate the beginning of a press.

Fig. 5 shows two ways of doing this. The first reductionconsiders the distance to the robot, denoted N , and the dis-tance to the caregiver, denoted CG, in Eq. 1.

N/(N + CG) (1)

The second reduction also considers the distance to the ex-perimenter, denoted E, in Eq. 2, so as to retain data on somecases in which participants sought the experimenter’s com-pany while interacting with the robot and not the caregiver.

N/(N + CG + E) (2)

In either case, we have reduced multiple series of variablemagnitude to a single, unit-less number scaled [0, 1], wherea 1 value means closest to caregiver and a 0 value meansclosest to the robot, regardless of the absolute distance be-tween child and robot or child and caregiver. We will returnto our first goal of representing a single child’s interactionwith a robot over time, but we move now to our second goal,to compare and contrast all participants with each other, withthese raw Euclidean distance data.

Recall that each interaction varies in length due to po-tential buffer time between presses in a single interaction.The buffer time, lasting up to one minute between pressesas needed, was included in the last press that occurred. Forexample, if Child A needed a 40 second break after a oneminute press, but Child B didn’t need a break and only usedtwo seconds after a one minute press, the press lasted for100 seconds for Child A but only 62 seconds for Child B.

This flexibility in interaction time naturally raises thequestion of how to compare these variable length data. Thetime difference between presses is capped at 60 seconds, andany time between presses is used to draw the participant’s at-tention back to the robot. By and large, the excess time wasspent by the participant getting a snack or toy, playing withother items in the room, or talking to their caregiver; noneof the buffer time was spent interacting with the robot whilethe robot was moving autonomously. Therefore, we considerthe time between presses to be noise.

To remove this noise, we choose a simple method of align-ing all presses in all interactions, and we truncate each pressto the length of the shortest occurrence of that press over allparticipants. For example, say Child A took 60 seconds dur-ing Press 1 with a 30 second break, then 180 seconds duringPress 2 with a 10 second break. Say Child B took 60 secondsduring Press 1 with a 2 second break, then 180 seconds dur-ing Press 2 with a 2 second break. It should be noted that inmost cases the experimenter manually starts the next pressfor attention with the robot, so short breaks of 1-3 seconds issimply the time taken to reach over and push buttons on therobot (or occasionally, to first re-orient the robot towards thechild if they shifted position). Our first data exploration onlycompares participant reactions to the robot while the robot isactively moving or speaking; thus, Press 1 is truncated to 62seconds for both participants and Press 2 is truncated to 182seconds for both participants. Alternatively, we could havetimed the presses off-line and truncated participant data withthose timings, but in practice, some participants progressedthrough the interactions in immediate succession and thisoff-line timing was unnecessary.

The effect of such data loss, i.e. 28 seconds after Press1 and 8 seconds after Press 2 in the above example, ad-mittedly contains some distance data. Either the participantdidn’t move between presses or approached the robot againfrom somewhere else in the room. In theory, a participantmight have outlasted the one minute buffer time and startedthe next press further from the robot than they were at the(truncated) end of the previous press. However, in practice,this did not happen; only the first two cases occurred. Thus,minimal interaction information was lost due to this trunca-tion. Comparisons of these extraneous portions of time thatmay include redirection, however, are left as future work.

Following this alignment and truncation procedure for ev-ery press, we next used several classic similarity measuresfor comparing interactions. With the shortened, three chan-nel time series, we applied four classic similarity measuresas given in (Cassisi et al. 2012): mean similarity, root meansquare similarity, peak similarity, and Pearson’s correlationfunction. In any case where a participant did not finish allseven presses, we compared only the presses he did finishwith other participants. The mean similarity and root meansquare similarity uses the similarity between two numbersas defined in Eq. 3.

numSim(x, y) = 1− |x− y||x|+ |y|

(3)

Let two time series, X = x1, x2, ..., xn and Y =y1, y2, ..., yn. Then we define mean similarity in Eq. 4, root

mean square similarity in Eq. 5, peak similarity in Eq. 6, andthe cross-correlation or Pearson’s correlation function in Eq.7. In all cases, n is the length of the time series X and Y ; inPearson’s correlation function, l allows a shifted comparisonof positions within the second time series.

tsim(X,Y ) =1

n

n∑i=1

numSim(xi, yi) (4)

rtsim(X,Y ) =

√√√√ 1

n

n∑i=1

numSim(xi, yi)2 (5)

psim(X,Y ) =1

n

n∑i=1

[1− |xi − yi|

2max(|xi|, |yi|)

](6)

rXY =

∑ni=1(xi − X)(yi−l − Y )√∑n

i=1(xi − X)2√∑n

i=1(yi−l − Y )2(7)

Figure 6: Mean similarity between the 59 children, with eachnumber represented as the darkness of a grey-scale pixel.White pixels indicate total similarity; black pixels representcomplete dissimilarity. This is only the distance to caregiver,sorted in order of average intensity of each child partici-pant’s similarity scores to the other participants.

We first considered each data stream individually, com-paring each child to every other child (using the shorterof the two data streams if a participant did not finish allpresses). The similarity measure for each child to every otherchild was multiplied by 255 and turned into a three-part tu-ple, which was then used as the grey-scale pixel color in a59x59 image where a row and column correspond to the sim-ilarity measure between two children. Fig. 6 shows the simi-larity measures between 59 participants using the distance to

caregiver as the only data stream, using the mean similaritymeasure. The participants have been ordered by the intensityof the average of the entire vector of differences per child.

Figure 7: Best viewed in color. Mean similarity between the59 children, in three color channels: caregiver in green, ex-perimenter in red, and robot in blue. White pixels indicatetotal similarity; black pixels represent complete dissimilar-ity. This represents all channels of the three-channel timeseries that form a participant’s proxemics, sorted in order ofaverage intensity of distance to caregiver of each child par-ticipant’s similarity scores to the other participants, resultingin the same sorting order as the single channel in Fig. 6.

Figure 8: The combinedmean similarity measurebetween the 32 male par-ticipants, sorted by averagedistance metric of thedistance to caregiver.

Figure 9: The combinedmean similarity measurebetween the 28 femaleparticipants, sorted by byaverage distance metric ofthe distance to caregiver.

Figures 7, 8, and 9 show combined similarity measuresbetween all 59 participants, just the 32 males, and just the28 females, respectively. These images use all three chan-nels, again using the mean similarity measure; the similarity

measure between the participants’ distance to robot was as-signed to the blue color channel, caregiver distance was as-signed green, and experimenter distance was assigned red.Each image has been ordered by the average intensity of thecaregiver channel of the time series, meaning the order ofparticipants in Fig. 6 and Fig. 7 are the same. Note that theimages look similar, but the variation in distance to robotand distance to experimenter adds further characterizationof the participants. In Figures 6 – 9, the order of the chil-dren is the same across the rows and columns, resulting in asymmetric image split from top left to bottom right, sortedby the summation of the intensity of differences per child. Inthese images, total similarity comes out as white, and totaldissimilarity comes out as black.

Thus far, we have used absolute and continuous distances,even in our distance metrics in Eq. 1–2. Consider that shorterdistances between a participant and another actor probablyindicate more comfort or interest, and longer distances in-dicate less comfort or interest. A participant sitting two feetaway from the robot and five feet away from the caregiverprobably indicates strong comfort with or interest in therobot (e.g. Fig. 10). However, if a participant is seated ontheir caregiver and the caregiver is seated two feet from therobot, they probably need comfort from their caregiver whilethey watch the robot (e.g. Fig. 13). Two feet of distance be-tween the participant and the robot can indicate very differ-ent levels of comfort and interest. We therefore must cre-ate a system which considers relative, rather than absolute,distances to quantify and discretize the levels of interest orcomfort the child has in the robot.

Figure 10: Zone 1 – the childis closest to the robot.

Figure 11: Zone 2 – the childis nearest the experimenter.

Figure 12: Zone 3 – the childis not close to anyone.

Figure 13: Zone 4 – the childis nearest the caregiver.

There are four basic places a child can be during an inter-

Figure 14: Percent of time spent in each zone for one par-ticipant for each press, from P1 to P7. Note the progressionfrom Zone 4, closest to caregiver, to Zone 1, closest to robot.This can indicate greater comfort with the robot or less needfor comfort from a caregiver, or both, as time progressed.

action: closest to their parent, closest to the robot, closest tothe experimenter, or not close to any of these. If the child isexactly mid-way between the robot and caregiver, we cannotsay there is a clear preference for one or the other. Therefore,we want to qualify definite spaces that indicate a preferencefor robot (or experimenter) or caregiver, with a buffer spacebetween them. The simplest method of doing so is to splitthe distance between the robot and the caregiver into thirds,giving the caregiver, robot, and the buffer an equal share ofthe distance. If the child is within the third of distance closestto the caregiver, we say they have a preference for the care-giver (Fig. 13); if the child is within a third of that distanceto the robot, we say they have a preference for the robot(Fig. 10). If they are much closer to the experimenter thanthe robot while within that distance (which happens very in-frequently), we say they have a preference for the experi-menter (Fig. 11). If none of those situations apply, the childis near no-one (Fig. 12). We call these locations ‘zones,’ andnumber them from 1 to 4, in order of preference from robot(Zone 1), experimenter (Zone 2), no-one (Zone 3), to care-giver (Zone 4). When averaging zones across time, we usethe mode zone during that time.

We first show a sample breakdown of percent time spent

in each zone over an entire interaction, separated by press, inFig. 14 (for the same participant as in Fig. 1, 3, and 4). Thisfigure shows the presses in order from 1 to 7, top to bot-tom, as the percent time in each press spent in each zone.This child spent most of the time in Press 1 and Press 2near their caregiver; in Press 3 and Press 4 they began toexplore the room and spent time away from (or in between)the robot and the caregiver. In Press 5 more time was spentaway (or between) the robot and caregiver, until Press 6 wasentirely spent nearest the robot or experimenter, and Press 7was more mixed. This concludes our first goal, representinga single child’s interaction proxemics.

Figure 15: Zones throughout every child’s interaction, sortedtop to bottom by increasing age. White indicates time closestto caregiver, and black indicates closest to robot.

Fig. 15 shows all children’s zones over time; the blue barindicates a lack of press information (i.e. that the experimenthad ended). White indicates the zone closest to the caregiver,light gray indicates not close to anyone, darker gray indi-cates closest to experimenter, and black indicates closest tothe robot. The 59 children here are sorted from youngestto oldest, with the youngest at the top. A border has beenincluded to make the top two rows more obvious – thesechildren were always in Zone 4, or nearest the caregiver. Atthis juncture, it appears that as age increases, the zones tendcloser towards the robot.

We show the same zone information sorted by age withjust the male participants in Fig. 16 and with just the femaleparticipants in Fig. 17. Note that the tendency towards beingcloser to the robot is more visible in boys. We also graphthe total percent of time per child in each zone as a threedimensional plot in Fig. 18. Note that there are two areas ofconcentration: 100% near caregiver, and 100% near robot. Athree dimensional plot of age (in months) versus preferencefor robot, experimenter, or caregiver (over the entire inter-action) is shown in Fig. 19. Note that the age range variesacross the cluster of individuals that spent 100% of their timenear the robot, but there is a slight effect of age on percenttime spent with robot or caregiver. This concludes our sec-ond goal of comparing and contrasting all participants.

5 Conclusions and Future WorkIn this paper we have shown several ways of collapsing ourchild-robot interaction proxemics data. We began by show-

Figure 16: Zones throughout the male participants’ interac-tion, sorted top to bottom by increasing age.

Figure 17: Zones throughout the female participants’ inter-actions, sorted top to bottom by increasing age.

Figure 18: Percent of time spent in each zone for all partic-ipants; some jitter has been added to all points to make theclusters at both ends more visible.

ing the simple Euclidean distance between the child par-ticipants and other actors of interest, and normalizing thatthree-channel time series into a single metric over the sameamount of time. We then employed similarity measures overthe time series both as three separate series and a three-channel series, condensing variable length interactions to asingle distance measure between participant pairs. We thendiscussed how using the space around the parent, robot, orexperimenter could be simplified into a single zone at anypoint in time, and showed how this can demonstrate progres-sion over time from one zone to another. We also succinctlyshowed all participants’ progressions over time.

Future work will include broader methods of participant

Figure 19: Percent of time spent in caregiver and robot zones(without any adjustment for time spent in no-one’s range),plotted against age in months. This combines robot and ex-perimenter zones. Note a slight preference for older childrenfor being closer to the robot.

comparison, participant clustering, and further validationsof statistical significance between male and female partic-ipants and over age. First, while our initial method of equal-izing participants’ time series is truncation of excess noise,it could be that the time lapse between presses holds keyinformation– for example, longer time between presses po-tentially indicates a less interested child or a more activechild that repeatedly needs their attention drawn back to therobot. Thus, other methods that include the entire interac-tion, such as dynamic time warping, on a press-by-press ba-sis or on the normalized nearness metric across all partici-pants, might show useful differences.

Second, we will explore how to group or cluster these par-ticipants; while straightforward methods such as k-means re-quire a priori knowledge of how many clusters to use, othermethods such as k-medoids may prove to give reasonable (ifnot optimal) solutions. We have also explored choosing clus-ters based on the visual data (e.g. the two green patches andthe pink or blue strips in Fig. 6), but we need to explore po-tential relationships between participants’ similarity scoresand participants’ location zones over the interactions. Lastly,while we have visually explored the sex and age differencesof participants, it is not yet clear how much statistical signif-icance the differences hold.

6 Acknowledgments

The authors thank the National Science Foundation“I/UCRC: Robots and Sensors for the Human Well-being,”[NSF IIP-1439728] and the MnDRIVE RSAM initiative atthe University of Minnesota for partial support for this work,and the anonymous reviewers for their feedback.

ReferencesAsada, K.; Tojo, Y.; Osanai, H.; Saito, A.; Hasegawa, T.; andKumagaya, S. 2016. Reduced personal space in individualswith autism spectrum disorder. PLoS One 11(1).Cassisi, C.; Montalto, P.; Aliotta, M.; Cannata, A.; and Pul-virenti, A. 2012. Similarity measures and dimensionality re-duction techniques for time series data mining. In Karahoca,A., ed., Advances in Data Mining Knowledge Discovery andApplications. InTech.Dautenhahn, K., and Werry, I. 2004. Towards interactiverobots in autism therapy: background, motivation and chal-lenges. Pragmat. Cogn. 12(1):1–35.Diehl, J. J.; Schmitt, L. M.; Villano, M.; and Crowell, C. R.2012. The clinical use of robots for individuals with autismspectrum disorder: a critical review. Res. Autism Spectr. Dis-ord. 6(1):249–262.Fasching, J.; Walczak, N.; Morellas, V.; and Papanikolopou-los, N. 2015. Classification of motor stereotypies in video.IEEE Int. Conf. Intell. Robot. Syst. 4894–4900.Feil-Seifer, D. J., and Mataric, M. J. 2011. Automated de-tection and classification of positive vs. negative robot inter-actions with children with autism using distance-based fea-tures. Proc. 6th Int. Conf. Human-robot Interact. - HRI ’11323–330.Hashemi, J.; Spina, T. V.; Tepper, M.; Esler, A.; Morellas, V.;Papanikolopoulos, N.; and Sapiro, G. 2012. Computer vi-sion tools for the non-invasive assessment of autism-relatedbehavioral markers. 2012 IEEE Int. Conf. Dev. Learn. Epi-genetic Robot. ICDL 2012 1–33.Hashemi, J.; Tepper, M.; Vallin Spina, T.; Esler, A.; Morel-las, V.; Papanikolopoulos, N.; Egger, H.; Dawson, G.; andSapiro, G. 2014. Computer vision tools for low-cost andnoninvasive measurement of autism-related behaviors in in-fants. Autism Res. Treat.Mead, R.; Atrash, A.; and Mataric, M. J. 2013. Automatedproxemic feature extraction and behavior recognition: ap-plications in human-robot interaction. Int. J. Soc. Robot.5(3):367–378.Pennisi, P.; Tonacci, A.; Tartarisco, G.; Billeci, L.; Ruta, L.;Gangemi, S.; and Pioggia, G. 2016. Autism and socialrobotics: A systematic review. Autism Res. 9(2):165–183.Scassellati, B.; Mataric, M. J.; and Admoni, H. 2012.Robots for use in autism research. Annu. Rev. Biomed. Eng.14(1):275–294.Scassellati, B. 2005. Quantitative metrics of social responsefor autism diagnosis. Proc. - IEEE Int. Work. Robot Hum.Interact. Commun. 2005:585–590.Scassellati, B. 2007. How social robots will help us to diag-nose, treat, and understand autism. Robot. Res. 552–563.Zwaigenbaum, L.; Bryson, S.; Rogers, T.; Roberts, W.;Brian, J.; and Szatmari, P. 2005. Behavioral manifestationsof autism in the first year of life. Int. J. Dev. Neurosci. 23(2-3SPEC. ISS.):143–152.